The One-Hour Model Did Not Beat the Spread

This week I gave a small model one hour and a fresh $1,000 PredArena strategy account. The rule was simple: use live public order books, keep costs low, trade often, and let the result be visible.

The model was deliberately small. It watched tight Kalshi books, learned from the last few price changes, and bought the side where momentum and book imbalance suggested the next move could cover the spread. It was not a language model making grand claims. It was a few weights updated online, which is often the right size for an hour-long test.

What Happened

It ended down $9.78 after closing the book. The account placed 152 orders over 60 minutes, including 76 exits. Total fees recorded by PredArena were $5.30.

Metric	Value
Starting balance	$1000.00
Final account balance	$990.22
Realized PnL	$-9.78
Orders	152
Round trips	76
Gross bought	$114.38
Best closed trade	YES KXXISUCCESSOR-45JAN01-LXI: $0.00
Worst closed trade	YES KXFUSION-40-JAN01: $-0.23

The Lesson

The hard part was not finding movement. Prediction markets move. The hard part was paying the toll to enter and leave. A one-cent spread plus taker fees can turn a clever-looking signal into a small leak.

That is a useful result. The next version should be less eager and more patient: post limits when the queue is favorable, trade fewer markets, and require a larger expected move unless the fee schedule is waived. In short, the model should make the fill part of the prediction, not an afterthought.

This run was not profitable. But it made the next test clearer, which is the other thing an honest trading experiment can earn.